Comment by q4y2u35uk
5 hours ago
People who think this do not understand LLMs. LLMs thrive in abstractions just like humans. Assembly code introduces far too much entropy, requires far too much context and documentation, and is far too easy to get horribly wrong. Even after generating large amounts of synthetic training data, it will never be practical.
But if the LLM can use decompilers in the process? It already uses many tools (grep, python, sed, etc...); so why not a decompiler and then take it from there?
I've seen this approach already in practice: LLM+decompiler give better code than the original code that the binary was compiled from.
I'd imagine it's just a lot less token efficient than having it write the code in a high level language.